Linkage mapping of bipolar affective disorder
To the Editor
Numerous studies have attempted to map genes for bipolar affective disorder (BP). Although support for linkage to putative loci has been reported by several groups, consistent replication is lacking [DeLisi & Crow, 1999]. The uneven findings have been ascribed to the complex inheritance of the disease, leading to reduced statistical power and difficulties in discerning a true from a false positive, or uncertainties in research procedures (Risch and Botstein, 1996; Baron, 1997).
Two recent genome scans with seemingly contradictory results have refocused attention on the complex task at hand. The first scan [Detera-Wadleigh et al., 1999] revealed possible linkage at several chromosomal sites, including putative loci reported independently by other investigators, on 4p16 [Blackwood et al., 1996], 12q23-24 [Ewald et al., 1998], 18q22-23 [Freimer et al., 1996], and 21q22 [Aita et al., 1999; Straub et al., 1994], i.e., a potential confirmation of previous findings. In contrast, the other scan [Friddle et al., 2000] found no linkage across the genome. (Some genomic regions showed modest LOD scores but the investigators viewed their results as largely negative.) The disparate results are noteworthy because the two studies employed similar protocols regarding ascertainment (unilineal multiplex pedigrees obtained from a U.S. general population), phenotype definition (narrow to broader diagnostic boundaries), and genetic models (parametric and nonparametric analyses). Also of interest was the fact that Fiddle et al. [2000] were unable to replicate their previous findings on 18p11 and 18q21-22 [McMahon et al., 1997; Stine et al., 1995], which were based on a subsample of their current pedigree set.
Friddle et al. [2000] attributed their generally negative results to the genetic complexity of BP, in particular nonallelic heterogeneity, whereby the disease is caused by rare dominant genes segregating in the population, or a complex inheritance model, with different, common disease alleles, which can jointly predispose to the illness. While these are plausible scenarios often invoked to account for variable linkage results in complex traits, there are complementary explanations:
SAMPLE STRATIFICATION
The dissection of complex traits such as BP into more homogeneous subsets has long been considered a useful means to detect an otherwise elusive linkage. In a previous analysis of a subset of their sample (28 of 50 families), some of the investigators in the Friddle et al. [2000] study stratified their pedigrees based on the type of parental inheritance [Stine et al., 1995]. They found significant linkage to 18q21-22 in pedigrees with paternal transmission (those in which the illness appeared to be transmitted from the father's side of the pedigree) using parametric analysis (LOD = 3.51), and suggestive linkage using nonparametric ASP analysis (P = 0.00002), both at D18S41. A similar pattern of results was shown for 18p11, though the linkage evidence was less pronounced. This intriguing finding has spawned numerous attempts at replication, with conflicting results [reviews: Van Broeckhoven & Verheyen, 1999; Baron & Knowles, 1999].
In their recent analysis of the overall data [Friddle et al., 2000], the investigators refrained from including this effect in the analysis. While acknowledging that such analysis would be reasonable, they also noted that it was not part of their original hypothesis, and that there are uncertainties about the inclusion of this variable. Uncertainties regarding the detection and origin of parent-of-origin effect have been described by other investigators, who raised concerns about sample artifacts or chance fluctuations in pedigree patterns [Baron, 1997; Kato et al., 1996; Knowles et al., 1998]. However, the previous positive findings in a subsample of this cohort [Stine et al., 1995], and the fact that other studies using genomewide methods, or focusing exclusively on chromosome 18, included this effect in the analysis [review: Van Broeckhoven & Verheyen, 1999], suggest that this approach may have a role in sorting out potential linkages (see False-Negative vs False-Positive Linkage for caveats).
MARKER DENSITY AND INFORMATIVENESS
The average marker density in most genome scans of BP, including the one by Friddle et al. [2000], is 10–12 cM. Although this marker spacing is assumed to capture linkages to loci residing in marker-to-marker intervals, not all regions of the genome are equally covered by equally informative markers, and genuine linkages can escape detection due to gaps in marker coverage. A case in point is our recent genome scan of BP [J. Liu, V.M. Aita, J. Terwilliger, T.C. Gilliam, and M. Baron, unpublished data] which, in spite of marker spacing averaging 10 cM over the entire genome, did not corroborate the evidence we reported elsewhere [Aita et al., 1999; Straub et al., 1994] for linkage to 21q22. As it turns out, the marker nearest to D21S1260, the locus with the highest LOD score in a subsequent study of the same pedigree set, with <2 cM marker specing on 21q22 [Aita et al., 1999], was 8 cM away. Evidently this distance was sufficiently large to prevent linkage detection in the genome scan. Interestingly, the average marker spacing in the Detera-Wadleigh et al. [1999] scan, which appeared to support some previously reported linkages, was denser than most other reported scans (about 6 cM).
In considering this lesson in linkage detection, we [Aita et al., 1999] observed the following: “This serves as an important caveat to the interpretation of results from genome scans—in particular, overconfidence in the marker coverage reported for linkage to complex diseases. Random fluctuations in marker informativeness in any given family across a pedigree series could result in errors in linkage detection.” Although our own experience need not necessarily apply to each and every genome scan reported for BP, it serves a cautionary notice. And while high-density (1–2 cM) genomewide scans may not be practicable, replication studies using dense marker coverage of specific candidate regions may be helpful in evaluating putative linkages.
FALSE-NEGATIVE VS. FALSE-POSITIVE LINKAGE
Although failure to replicate linkage can be attributed to various factors including misspecified models, low statistical power to uncover genes of modest effect, and gaps in marker spacing, the possibility that some of the published linkages are false-positive is an alternative explanation. For example, Friddle et al. [2000] mention 18p as one of the most widely studied regions, where Berrettini et al. [1994] first published evidence for linkage to BP. They note that subsequent studies have supported this finding, including a meta-analysis on several data sets [Lin & Bale, 1997]. However, the pattern of results is far from uniform. As reviewed [Baron & Knowles, 1999, 2000; Van Broeckhoven & Verheyen, 1999], most studies were nonsupportive or negative, including samples with adequate statistical power to detect the presumed linkage. Also, most of the negative studies were not included in the meta-analysis, which was confined to a single locus (D18S37), with no multipoint results on nearby markers. Other analyses of the same data set showed little or no evidence of linkage [Daly et al., 1997; Davis et al., 1997; Durner & Abreu, 1997; Li & Schaid, 1997]. This raises the possibility that the some of the positive linkage data were spurious or overstated.
Conflicting results have been noted for other putative linkages for BP [DeLisi & Crow, 1999]. Aside from chance statistical fluctuations, which can be more consequential in small samples, uncertainties in research procedures such as multiple test effects are a likely source of false positives. Such effects may result from the proliferation of analytical permutations involving genetic models, disease phenotypes, and subsets of a large data set. Such hazards are not confined to psychiatric genetics. A case in point is a widely discussed report on linkage between type 2 diabetes mellitus and chromosome 12 near D12S1349 [Mahtani et al., 1996]. The investigators found no significant evidence of linkage when the families were analyzed together, but strong evidence for linkage when families were classified according to insulin levels in affecteds. However, with subsequent correction for family selection, the linkage was no longer statistically significant [Kong et al., 1997]. Whether or not this example applies to the difficulties in replicating linkage of BP to 18p and 18q in samples stratified according to parent-of-origin effect remains to be seen. However, the use of proper correction factors to circumvent multiple test effects should be encouraged.
The recent genome scans underscore the uncertainties in linkage results for BP. Large-scale studies with standardized methods, power to detect modest gene effects, and adequate marker density across the genome, will be needed to sort out true linkages portending BP loci.
